Hollywood writers and directors are taking $12–$200/hour gigs to train AI to do their jobs.
What happened: The Guardian reports that experienced film and TV creatives are signing up with AI training agencies that contract with companies including Anthropic and OpenAI, teaching models how to write screenplays, build pitch decks, and plan shoots. Pay cited ranges from $12 to $200 an hour. Screenwriter-producer Ruth Fowler, whose credits include BBC One’s Rules of the Game and Paramount+’s Little Disasters, said she started because she was “always broke” and that production is down by about 35%. One task asked her to schedule a hypothetical two-day shoot from emails, a script, permits, daylight, hazards, and child-protection needs. An anonymous LA documentary director reviewed muddy Little League audio to teach transcription and speaker IDs, then told friends he had been “handed a shovel and asked to dig the grave of my profession.” FilmLA Research found Los Angeles shoot days fell 48% from 2021 to 2025. Bureau of Labor Statistics data cited in the piece put U.S. motion-picture and sound-recording jobs at 326,000 in May 2026, down 28% from 450,000 in July 2022. Micro1 was advertising for producers with more than five years’ credited experience at up to $85 an hour to design production-management evaluation tasks. USC instructor Jody Wheeler said models churn ideas but are “not going to be generating Oscar-winning scripts any time soon.”
Why it matters: The labor market for film craft is now paying practitioners to encode their own replacement tasks while industry employment is already down. The measurable record is training-gig volume versus returning union work, whether agencies keep needing senior humans, and whether shoot-day and BLS employment recover independently of the AI pipeline.
Apple cuts more than 200 roles across Vision Pro, Siri, and on-device AI teams.
What happened: TechCrunch, citing Bloomberg, reports that Apple is eliminating more than 200 positions: about 100 on the Vision Pro headset team and about 100 across Siri and Intelligent Systems Experience, the group that integrates AI into devices. Sources described a reshuffle toward newer AI efforts and new devices. Apple confirmed the cuts, saying it was looking “to evolve our business to deliver the best experiences for our users” and that “while we will create new roles as part of this change, it will also impact a limited number of existing roles.” The same week’s context is Apple’s attempt to recover its AI position, higher Mac and iPad prices after an AI-driven memory shortage, and a lawsuit accusing OpenAI of trade-secret theft. The report does not publish a company-wide headcount percentage.
Why it matters: Even firms late to generative AI are using the AI transition as the rationale for cutting yesterday’s hardware and assistant teams. The measurable record is whether “new roles” offset the 200-plus cuts, Siri quality after the shrink, and whether Vision Pro investment keeps falling.
ERCOT says it will audit hundreds of Texas data-center proposals by December 10.
What happened: Utility Dive reports that the Electric Reliability Council of Texas intends to finish an audit of hundreds of data-center proposals by Dec. 10 so its Batch Zero large-load study can continue and interconnections can resume. Gov. Greg Abbott on Aug. 3 called for a pause on new data-center interconnections until questions about energy, water, and public funds are answered. Officials told the Public Utility Commission of Texas that about 300 data centers of 75 MW or larger are in Batch Zero, and that ERCOT will also run a community-impact review on data centers and crypto facilities of 25 MW and above. Abbott’s letter said interconnection-queue requests total about 474 GW, roughly 90% data centers — more than five times ERCOT’s record peak demand, with a substantial share possibly speculative or duplicative. Counsel Chad Seely said RFIs to provisionally qualified large loads should start by late August and early September, with a comprehensive verification report due to the PUCT a week before the December meeting. The Aug. 7 Batch Zero classification deadline already slipped; Seely said the April 9, 2027, interconnection-study date will not hold. Separately, 17 large loads that already passed stability assessment — mostly data centers, about 6.6 GW peak ramping over about five years — still need the audit before energization. Utilities were also asked to update an April snapshot of 157 medium data-center and crypto facilities totaling almost 8,800 MW.
Why it matters: Texas is treating the AI-era queue as something to verify, not automatically interconnect. The measurable record is how much of the 474 GW survives the audit, whether the 6.6 GW already at the last gate gets energized, and how far the study calendar slips.
PJM takes a second swing at surplus interconnection after almost no projects used the first one.
What happened: After a 2025 reform produced little, PJM is again trying to let new generation and storage plug into leftover interconnection rights at existing plants. Surplus interconnection service can be faster and cheaper than a full queue study because it uses an existing point of interconnection. Since 2023 PJM has received eight surplus applications and approved two, according to a November staff update. By comparison, an advisory-firm tally as of June 30 put surplus requests at 14.8 GW in MISO and 14.3 GW in SPP; PacifiCorp was reviewing 33 projects totaling 5.2 GW as of Aug. 13. A 2025 UC Berkeley working paper estimated PJM thermal and renewable plants could host about 150 GW of surplus solar, wind, and storage, though lost tax credits have reduced that potential. PJM missed reserve-margin targets in its last two capacity auctions, with shortfalls of about 6.5 GW for 2027/28 and 6.8 GW for 2028/29. Staff this month floated letting hybrid resources bid capacity as one unit while settling energy and ancillaries separately — a fix developers say is required for batteries added to solar farms. Indiana and Virginia this year directed utilities to study surplus interconnection on their systems.
Why it matters: Large-load growth is colliding with a queue that cannot add capacity fast enough. The measurable record is whether PJM’s rule tweak produces operating surplus megawatts, not just issue charges, and whether other RTOs stay ahead.
A California fight over “Italian brainrot” memes may decide who owns AI-generated characters.
What happened: NPR reports that Do Big Studios, maker of the hit Roblox game Steal a Brainrot, is in federal court against French startup Mementum Lab over characters such as Tung Tung Sahur. Do Big’s amended July complaint says the characters cannot be copyrighted because “copyright protection requires human authorship, and AI-generated material does not qualify.” Mementum represents young creators and argues they should share in what it calls potentially hundreds of millions of dollars if the memes become franchises. Court records identify Tung Tung Sahur’s creator as Indonesian artist Fernanda Bagas Indrastata (Noxa), who made the kentongan-drum character in about 15 minutes with seven prompts. Mementum has countersued for trademark infringement rather than registering a U.S. copyright, a move Do Big’s lawyer says is an attempt to dodge Copyright Office guidance that “simply typing prompts into an image generator is not authorship.” UCLA’s Mark McKenna notes the Office’s 2025 report left AI-assisted works to a case-by-case human-authorship test, and says originality thresholds are generally low even if trademark is the wrong vehicle.
Why it matters: Kids’ meme economies are becoming the test case for whether a handful of prompts counts as authorship. The measurable record is whether the court reaches copyright at all, how it treats seven-prompt characters, and whether trademark becomes the back door for AI art.
A Nature Health perspective: consumer chatbots are being wired into care pathways, not just searched for symptoms.
What happened: In Nature Health, Yilan Wu, T. Y. Alvin Liu, and Pearse A. Keane argue that recent consumer-AI products are embedding themselves in care pathways by linking to personal health records and issuing care recommendations — and that public-health systems need an explicit regulatory response. The perspective sits in a 2026 landscape that already includes OpenAI’s ChatGPT Health, Perplexity Health’s records-and-wearables pitch, Amazon Health AI with One Medical, Microsoft Copilot for Health, and Ant Group’s AQ app, which South China Morning Post coverage cited in the paper’s references put above 100 million users. The authors treat this as an infrastructure shift: the “front door” of care is becoming an algorithm that sits outside traditional clinic software. The freely available extract is a perspective, not a trial; it does not publish new diagnostic accuracy or outcome data of its own.
Why it matters: Clinical AI policy written for hospital EHR tools will miss consumer apps that already sit between patients and clinicians. The measurable record is whether regulators classify record-linked chatbots as care, what evidence they require, and whether health systems can see what patients were told before they arrive.
A UC Berkeley math professor used AI to edit an op-ed attacking students’ missing math skills.
What happened: The Guardian reports that Zvezdelina Stankova, a UC Berkeley math professor, admitted using AI to help edit a roughly 2,000-word San Francisco Standard op-ed that said some students were “five to eight years” behind and lacked middle-school fractions and algebra, blaming the UC system’s test-blind admissions. Daily Californian journalists ran the piece through Pangram, which flagged about 33% as AI-generated or AI-assisted. Stankova said the article reflected “several hundred person-hours of intensive human work, of which about 80 hours are my own,” and that AI helped locate documents while “all analysis is the result of the team members.” The Standard said humans must stand behind every article. Stankova and more than 3,000 other UC faculty signed a June letter backing admissions testing; she said AI also helped edit that letter. Her analysis claimed the share of students with a “severe deficit” for calculus I tripled after test-blind admissions. The UC academic senate said in late July it would review whether standardized tests return, affecting fall 2028 admissions at earliest. AI-council member Camille Crittenden said most detection tools “are still quite unreliable and lack nuance”; Pangram claims a 99.66% identification rate.
Why it matters: Faculty AI use is now colliding with the same integrity debate aimed at students, inside the fight over who gets into elite STEM classrooms. The measurable record is whether UC restores testing, how op-ed desks disclose model edits, and whether detection scores change anyone’s mind about the underlying placement data.
